/torch_object_rnn

detection network with conv-rnn

Primary LanguagePython

torch_det_rnn

Artisanal Research for better temporal detectors

Features:

  • NEW!! very fast encode (~1ms) & decode boxes (14 ms) module (look at anchors.py)

My previous code was based on https://github.com/kuangliu/pytorch-retinanet/blob/master/encoder.py but was much slower (~300ms ~250ms by batch) as it was really sequential.

  • a custom moving mnist
  • a custom moving coco (applying an continuous homography to the photos)
  • ap metric measure
  • recurrent retina-net
  • reports with tensorboard

Coming Soon:

  • MOT metrics
  • COCO metrics
  • FoveaBox Encoding/ Decoding
  • Make "moving datasets" multithreaded using new Pytorch's IterableDataset